Output contract
Metric blueprint
- decision owner, question/action, and cadence;
- definition name/description and datasource identity;
- measure, aggregation, time dimension, grain, running-total behavior, and fixed filters;
- granularities, comparisons, favorable direction, number format, and optional goal;
- allowed dimensions with a one-line reason for each;
- insight settings with applicability rationale and entitlement/schema caveats;
- row-level settings only when a safe natural entity exists;
- justified filtered variants;
- evidence status, blockers, and implementation checklist.
Use exact enum/API values only when confirmed by the current tool schema; otherwise pair a plain-language recommendation with verify in target environment.
Metric program
Rank recommendations by decision value, definitional clarity, data readiness, explainable dimensions, and overlap with existing definitions. Default to the top five for readability, but honor a request for more. Include “not recommended” candidates only when the exclusion prevents a likely mistake.
Audit
Lead with scope and coverage: assessed, partial, failed, and unassessed properties. Then show composite and per-definition scores, highest-severity findings, observations, and concrete UI/API-ready changes. A score is not a claim that the underlying data is correct.
Optimization
Use columns for property, current, recommended, evidence, benefit/tradeoff, and change class (required, optional, blocked). Keep IDs internal unless needed for disambiguation or implementation.
Final checks
- Every field reference is observed or labeled unverified.
- Aggregation, grain, time behavior, and favorable direction are explicit.
- No goal, cardinality, insight availability, entitlement, or metric value is invented.
- Existing overlap and audit coverage are represented honestly.
- No mutation is implied unless it actually ran under the user’s authorization.